Skip to content

SoK: Private transformer-based model inference

Authors: Y. Chen; T. Lu; Z. Tang; B. Zhang; Z. Shi; Y. Luan; Z. Wang
Year: 2026

type: survey collection: sok-references status: auto

Links: ePrint 2026/491

Summary

Auto-extracted summary (unreviewed)

This SoK systematically reviews more than 61 private transformer-based model inference frameworks from 2022-2026, following a top-down analysis from outsourced deployment models through the cryptographic techniques used to concrete computation, communication and accuracy trade-offs, and identifies their limitations and research gaps. The authors also assess the reproducibility of prior systems and re-benchmark representative solutions under standardized configurations, deriving guidelines for choosing protocols in different deployment settings; a living tracker repository is provided.

Abstract

The growing demand for privacy-preserving Transformer inference has led to the emergence of numerous protocols designed to protect sensitive data and model parameters. These protocols utilize diverse cryptographic tools under varying assumptions, each presenting unique characteristics and tradeoffs between computation, communication, and accuracy. In this paper, we conduct a systematic and in-depth analysis of existing approaches from diverse performance perspectives, identifying their limitations and research gaps. We further evaluate the reproducibility of prior systems and rebenchmark representative solutions under standardized configurations. Our results yield a principled guideline for balancing protocol trade-offs under different deployment settings.
Source: pdf

Extracted details

Field Value Source Confidence
contribution Systematic in-depth analysis of 61+ private transformer inference frameworks across performance perspectives llm 0.70
contribution Top-down taxonomy from deployment models to cryptographic techniques to performance comparison llm 0.70
contribution Reproducibility evaluation of prior systems llm 0.70
contribution Re-benchmarking of representative solutions under standardized configurations llm 0.70
contribution Principled guideline for balancing protocol trade-offs under different deployment settings llm 0.70
contribution Public repository to continuously track new developments llm 0.70
tasks private transformer/LLM inference llm 0.70
tasks private text generation llm 0.70
tasks secure biomedical image analysis with transformers llm 0.70
code url https://zenodo.org/records/20319561 (URL found in the full text) regex 0.50
code url https://github.com/zama-ai/tfhe-rs (URL found in the full text) regex 0.50
code url https://github.com/tuneinsight/lattigo (URL found in the full text) regex 0.50
code url https://anonymous.4open.science/r/SoK4PPLLM llm 0.70
limitations Several open-source codes of surveyed systems could not be executed successfully, so reported numbers are used for those llm 0.70
llm relevance high llm 0.70
key result frameworks reviewed: 61+ (private transformer-based model inference, 2022-2026) llm 0.70

SoK reference [28]

Y. Chen, T. Lu, Z. Tang, B. Zhang, Z. Shi, Y. Luan, and Z. Wang, “SoK: Private transformer-based model inference,” Cryptology ePrint Archive, Paper 2026/491, 2026. [Online]. Available: https://eprint.iacr.org/2026/491

Listed on the reference page of the site.

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: MPC 121 [p.2], Secret sharing 24 [p.4], CKKS 19 [p.4], Function secret sharing 13 [p.2], BFV 12 [p.4], TFHE 10 [p.4], BGV 9 [p.4], Oblivious transfer 3 [p.4], TEE / confidential computing 3 [p.2], Garbled circuits 2 [p.4]
  • library: CrypTen 17 [p.12], Microsoft SEAL 14 [p.24], EzPC / CrypTFlow 13 [p.12], SecretFlow-SPU 10 [p.12], ABY / ABY3 6 [p.12], OpenFHE 4 [p.24], Lattigo 4 [p.23], TFHE-rs 4 [p.23], Pyfhel 4 [p.23], EVA / CHET 4 [p.11], MP-SPDZ 4 [p.12], HElib 3 [p.23], Orion 3 [p.11], HEaaN (library) 2 [p.26], HEonGPU 2 [p.23], HEIR 2 [p.11], Fhelipe 2 [p.11], Zama 1 [p.23], Liberate.FHE 1 [p.26], nGraph-HE 1 [p.11], PyTorch 1 [p.3], TensorFlow 1 [p.3], HuggingFace 1 [p.3]
  • model: BERT 96 [p.2], Transformer 34 [p.1], GPT 28 [p.3], ViT 22 [p.7], LLaMA 17 [p.3], LoRA 3 [p.6], ResNet 1 [p.12], RNN / LSTM 1 [p.21]
  • dataset: GLUE tasks (QNLI/RTE/MRPC/CoLA/QQP/MNLI/STS-B) 117 [p.13], SST-2 37 [p.27], CIFAR-10 6 [p.27], CIFAR-100 5 [p.27], ImageNet 5 [p.27], ARC 3 [p.28], WikiText 2 [p.27], SVHN 1 [p.27], IMDB 1 [p.29], Yelp 1 [p.29]
  • hardware: GPU 49 [p.2], FPGA 2 [p.13], ASIC 2 [p.13], Intel Xeon 2 [p.24]
  • technique: Semi-honest model 66 [p.4], Softmax 29 [p.3], GELU 18 [p.3], Rotations 18 [p.4], Compiler 18 [p.3], Two-party computation 16 [p.1], Quantization 15 [p.2], LayerNorm 14 [p.3], Bootstrapping 13 [p.4], Attention 10 [p.3], Malicious model 10 [p.4], ReLU 7 [p.3], SIMD packing 7 [p.4], Matrix multiplication 7 [p.3], Fine-tuning 7 [p.6], Lookup table 7 [p.7], Three-party computation 7 [p.3], Polynomial approximation 4 [p.2], Non-interactive 4 [p.15], KV cache 3 [p.9], Rescaling 2 [p.4], Modulus switching 2 [p.8], Convolution 2 [p.10], Knowledge distillation 2 [p.6]
  • task: Language modeling / generation 5 [p.1], Medical / healthcare 3 [p.1]
  • topic: Homomorphic encryption 123 [p.2], Private / secure inference 35 [p.2], Large language models 25 [p.6], Open source 25 [p.13], Privacy-preserving machine learning 19 [p.2], Benchmark 3 [p.1]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/sok-references/028_sok-private-transformer-based-model-inference.pdf sok-references 31 740.3 KB dbb2f676c45d

Database key: sok-private-transformer-based-model-inference · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set sok-private-transformer-based-model-inference COLUMN VALUE